lustre/obdclass/lu_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/lu_object.c
  37  *
  38  * Lustre Object.
  39  * These are the only exported functions, they provide some generic
  40  * infrastructure for managing object devices
  41  *
  42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_CLASS
  46
  47 #include <libcfs/libcfs.h>
  48 #include <linux/module.h>
  49 #include <libcfs/libcfs_hash.h> /* hash_long() */
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <lustre_disk.h>
  53 #include <lustre_fid.h>
  54 #include <lu_object.h>
  55 #include <lu_ref.h>
  56 #include <libcfs/list.h>
  57
  58 enum {
  59         LU_CACHE_PERCENT_MAX     = 50,
  60         LU_CACHE_PERCENT_DEFAULT = 20
  61 };
  62
  63 #define LU_CACHE_NR_MAX_ADJUST          128
  64 #define LU_CACHE_NR_UNLIMITED           -1
  65 #define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
  66 #define LU_CACHE_NR_LDISKFS_LIMIT       LU_CACHE_NR_UNLIMITED
  67 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
  68 #define LU_CACHE_NR_ZFS_LIMIT           10240
  69
  70 #define LU_SITE_BITS_MIN    12
  71 #define LU_SITE_BITS_MAX    24
  72 /**
  73  * total 256 buckets, we don't want too many buckets because:
  74  * - consume too much memory
  75  * - avoid unbalanced LRU list
  76  */
  77 #define LU_SITE_BKT_BITS    8
  78
  79
  80 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
  81 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
  82                 "Percentage of memory to be used as lu_object cache");
  83
  84 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
  85 CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
  86                 "Maximum number of objects in lu_object cache");
  87
  88 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  89
  90 /**
  91  * Decrease reference counter on object. If last reference is freed, return
  92  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  93  * case, free object immediately.
  94  */
  95 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  96 {
  97         struct lu_site_bkt_data *bkt;
  98         struct lu_object_header *top;
  99         struct lu_site          *site;
 100         struct lu_object        *orig;
 101         cfs_hash_bd_t            bd;
 102         const struct lu_fid     *fid;
 103
 104         top  = o->lo_header;
 105         site = o->lo_dev->ld_site;
 106         orig = o;
 107
 108         /*
 109          * till we have full fids-on-OST implemented anonymous objects
 110          * are possible in OSP. such an object isn't listed in the site
 111          * so we should not remove it from the site.
 112          */
 113         fid = lu_object_fid(o);
 114         if (fid_is_zero(fid)) {
 115                 LASSERT(top->loh_hash.next == NULL
 116                         && top->loh_hash.pprev == NULL);
 117                 LASSERT(list_empty(&top->loh_lru));
 118                 if (!atomic_dec_and_test(&top->loh_ref))
 119                         return;
 120                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 121                         if (o->lo_ops->loo_object_release != NULL)
 122                                 o->lo_ops->loo_object_release(env, o);
 123                 }
 124                 lu_object_free(env, orig);
 125                 return;
 126         }
 127
 128         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
 129         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 130
 131         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
 132                 if (lu_object_is_dying(top)) {
 133
 134                         /*
 135                          * somebody may be waiting for this, currently only
 136                          * used for cl_object, see cl_object_put_last().
 137                          */
 138                         wake_up_all(&bkt->lsb_marche_funebre);
 139                 }
 140                 return;
 141         }
 142
 143         /*
 144          * When last reference is released, iterate over object
 145          * layers, and notify them that object is no longer busy.
 146          */
 147         list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 148                 if (o->lo_ops->loo_object_release != NULL)
 149                         o->lo_ops->loo_object_release(env, o);
 150         }
 151
 152         if (!lu_object_is_dying(top) &&
 153             (lu_object_exists(orig) || lu_object_is_cl(orig))) {
 154                 LASSERT(list_empty(&top->loh_lru));
 155                 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 156                 bkt->lsb_lru_len++;
 157                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 158                 return;
 159         }
 160
 161         /*
 162          * If object is dying (will not be cached), removed it
 163          * from hash table and LRU.
 164          *
 165          * This is done with hash table and LRU lists locked. As the only
 166          * way to acquire first reference to previously unreferenced
 167          * object is through hash-table lookup (lu_object_find()),
 168          * or LRU scanning (lu_site_purge()), that are done under hash-table
 169          * and LRU lock, no race with concurrent object lookup is possible
 170          * and we can safely destroy object below.
 171          */
 172         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
 173                 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 174         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 175         /*
 176          * Object was already removed from hash and lru above, can
 177          * kill it.
 178          */
 179         lu_object_free(env, orig);
 180 }
 181 EXPORT_SYMBOL(lu_object_put);
 182
 183 /**
 184  * Put object and don't keep in cache. This is temporary solution for
 185  * multi-site objects when its layering is not constant.
 186  */
 187 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
 188 {
 189         set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
 190         return lu_object_put(env, o);
 191 }
 192 EXPORT_SYMBOL(lu_object_put_nocache);
 193
 194 /**
 195  * Kill the object and take it out of LRU cache.
 196  * Currently used by client code for layout change.
 197  */
 198 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
 199 {
 200         struct lu_object_header *top;
 201
 202         top = o->lo_header;
 203         set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
 204         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
 205                 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
 206                 cfs_hash_bd_t bd;
 207
 208                 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
 209                 if (!list_empty(&top->loh_lru)) {
 210                         struct lu_site_bkt_data *bkt;
 211
 212                         list_del_init(&top->loh_lru);
 213                         bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
 214                         bkt->lsb_lru_len--;
 215                 }
 216                 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
 217                 cfs_hash_bd_unlock(obj_hash, &bd, 1);
 218         }
 219 }
 220 EXPORT_SYMBOL(lu_object_unhash);
 221
 222 /**
 223  * Allocate new object.
 224  *
 225  * This follows object creation protocol, described in the comment within
 226  * struct lu_device_operations definition.
 227  */
 228 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 229                                          struct lu_device *dev,
 230                                          const struct lu_fid *f,
 231                                          const struct lu_object_conf *conf)
 232 {
 233         struct lu_object *scan;
 234         struct lu_object *top;
 235         struct list_head *layers;
 236         unsigned int init_mask = 0;
 237         unsigned int init_flag;
 238         int clean;
 239         int result;
 240         ENTRY;
 241
 242         /*
 243          * Create top-level object slice. This will also create
 244          * lu_object_header.
 245          */
 246         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 247         if (top == NULL)
 248                 RETURN(ERR_PTR(-ENOMEM));
 249         if (IS_ERR(top))
 250                 RETURN(top);
 251         /*
 252          * This is the only place where object fid is assigned. It's constant
 253          * after this point.
 254          */
 255         top->lo_header->loh_fid = *f;
 256         layers = &top->lo_header->loh_layers;
 257
 258         do {
 259                 /*
 260                  * Call ->loo_object_init() repeatedly, until no more new
 261                  * object slices are created.
 262                  */
 263                 clean = 1;
 264                 init_flag = 1;
 265                 list_for_each_entry(scan, layers, lo_linkage) {
 266                         if (init_mask & init_flag)
 267                                 goto next;
 268                         clean = 0;
 269                         scan->lo_header = top->lo_header;
 270                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 271                         if (result != 0) {
 272                                 lu_object_free(env, top);
 273                                 RETURN(ERR_PTR(result));
 274                         }
 275                         init_mask |= init_flag;
 276 next:
 277                         init_flag <<= 1;
 278                 }
 279         } while (!clean);
 280
 281         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 282                 if (scan->lo_ops->loo_object_start != NULL) {
 283                         result = scan->lo_ops->loo_object_start(env, scan);
 284                         if (result != 0) {
 285                                 lu_object_free(env, top);
 286                                 RETURN(ERR_PTR(result));
 287                         }
 288                 }
 289         }
 290
 291         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 292         RETURN(top);
 293 }
 294
 295 /**
 296  * Free an object.
 297  */
 298 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 299 {
 300         struct lu_site_bkt_data *bkt;
 301         struct lu_site          *site;
 302         struct lu_object        *scan;
 303         struct list_head        *layers;
 304         struct list_head         splice;
 305
 306         site   = o->lo_dev->ld_site;
 307         layers = &o->lo_header->loh_layers;
 308         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 309         /*
 310          * First call ->loo_object_delete() method to release all resources.
 311          */
 312         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 313                 if (scan->lo_ops->loo_object_delete != NULL)
 314                         scan->lo_ops->loo_object_delete(env, scan);
 315         }
 316
 317         /*
 318          * Then, splice object layers into stand-alone list, and call
 319          * ->loo_object_free() on all layers to free memory. Splice is
 320          * necessary, because lu_object_header is freed together with the
 321          * top-level slice.
 322          */
 323         INIT_LIST_HEAD(&splice);
 324         list_splice_init(layers, &splice);
 325         while (!list_empty(&splice)) {
 326                 /*
 327                  * Free layers in bottom-to-top order, so that object header
 328                  * lives as long as possible and ->loo_object_free() methods
 329                  * can look at its contents.
 330                  */
 331                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 332                 list_del_init(&o->lo_linkage);
 333                 LASSERT(o->lo_ops->loo_object_free != NULL);
 334                 o->lo_ops->loo_object_free(env, o);
 335         }
 336
 337         if (waitqueue_active(&bkt->lsb_marche_funebre))
 338                 wake_up_all(&bkt->lsb_marche_funebre);
 339 }
 340
 341 /**
 342  * Free \a nr objects from the cold end of the site LRU list.
 343  */
 344 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 345 {
 346         struct lu_object_header *h;
 347         struct lu_object_header *temp;
 348         struct lu_site_bkt_data *bkt;
 349         cfs_hash_bd_t            bd;
 350         cfs_hash_bd_t            bd2;
 351         struct list_head         dispose;
 352         int                      did_sth;
 353         unsigned int             start;
 354         int                      count;
 355         int                      bnr;
 356         unsigned int             i;
 357
 358         if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
 359                 RETURN(0);
 360
 361         INIT_LIST_HEAD(&dispose);
 362         /*
 363          * Under LRU list lock, scan LRU list and move unreferenced objects to
 364          * the dispose list, removing them from LRU and hash table.
 365          */
 366         start = s->ls_purge_start;
 367         bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 368  again:
 369         /*
 370          * It doesn't make any sense to make purge threads parallel, that can
 371          * only bring troubles to us. See LU-5331.
 372          */
 373         mutex_lock(&s->ls_purge_mutex);
 374         did_sth = 0;
 375         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 376                 if (i < start)
 377                         continue;
 378                 count = bnr;
 379                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 380                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 381
 382                 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 383                         LASSERT(atomic_read(&h->loh_ref) == 0);
 384
 385                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 386                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 387
 388                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 389                                                &bd2, &h->loh_hash);
 390                         list_move(&h->loh_lru, &dispose);
 391                         bkt->lsb_lru_len--;
 392                         if (did_sth == 0)
 393                                 did_sth = 1;
 394
 395                         if (nr != ~0 && --nr == 0)
 396                                 break;
 397
 398                         if (count > 0 && --count == 0)
 399                                 break;
 400
 401                 }
 402                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 403                 cond_resched();
 404                 /*
 405                  * Free everything on the dispose list. This is safe against
 406                  * races due to the reasons described in lu_object_put().
 407                  */
 408                 while (!list_empty(&dispose)) {
 409                         h = container_of0(dispose.next,
 410                                           struct lu_object_header, loh_lru);
 411                         list_del_init(&h->loh_lru);
 412                         lu_object_free(env, lu_object_top(h));
 413                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 414                 }
 415
 416                 if (nr == 0)
 417                         break;
 418         }
 419         mutex_unlock(&s->ls_purge_mutex);
 420
 421         if (nr != 0 && did_sth && start != 0) {
 422                 start = 0; /* restart from the first bucket */
 423                 goto again;
 424         }
 425         /* race on s->ls_purge_start, but nobody cares */
 426         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 427
 428         return nr;
 429 }
 430 EXPORT_SYMBOL(lu_site_purge);
 431
 432 /*
 433  * Object printing.
 434  *
 435  * Code below has to jump through certain loops to output object description
 436  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 437  * composes object description from strings that are parts of _lines_ of
 438  * output (i.e., strings that are not terminated by newline). This doesn't fit
 439  * very well into libcfs_debug_msg() interface that assumes that each message
 440  * supplied to it is a self-contained output line.
 441  *
 442  * To work around this, strings are collected in a temporary buffer
 443  * (implemented as a value of lu_cdebug_key key), until terminating newline
 444  * character is detected.
 445  *
 446  */
 447
 448 enum {
 449         /**
 450          * Maximal line size.
 451          *
 452          * XXX overflow is not handled correctly.
 453          */
 454         LU_CDEBUG_LINE = 512
 455 };
 456
 457 struct lu_cdebug_data {
 458         /**
 459          * Temporary buffer.
 460          */
 461         char lck_area[LU_CDEBUG_LINE];
 462 };
 463
 464 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 465 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 466
 467 /**
 468  * Key, holding temporary buffer. This key is registered very early by
 469  * lu_global_init().
 470  */
 471 static struct lu_context_key lu_global_key = {
 472         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 473                     LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
 474         .lct_init = lu_global_key_init,
 475         .lct_fini = lu_global_key_fini
 476 };
 477
 478 /**
 479  * Printer function emitting messages through libcfs_debug_msg().
 480  */
 481 int lu_cdebug_printer(const struct lu_env *env,
 482                       void *cookie, const char *format, ...)
 483 {
 484         struct libcfs_debug_msg_data *msgdata = cookie;
 485         struct lu_cdebug_data        *key;
 486         int used;
 487         int complete;
 488         va_list args;
 489
 490         va_start(args, format);
 491
 492         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 493         LASSERT(key != NULL);
 494
 495         used = strlen(key->lck_area);
 496         complete = format[strlen(format) - 1] == '\n';
 497         /*
 498          * Append new chunk to the buffer.
 499          */
 500         vsnprintf(key->lck_area + used,
 501                   ARRAY_SIZE(key->lck_area) - used, format, args);
 502         if (complete) {
 503                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 504                         libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
 505                 key->lck_area[0] = 0;
 506         }
 507         va_end(args);
 508         return 0;
 509 }
 510 EXPORT_SYMBOL(lu_cdebug_printer);
 511
 512 /**
 513  * Print object header.
 514  */
 515 void lu_object_header_print(const struct lu_env *env, void *cookie,
 516                             lu_printer_t printer,
 517                             const struct lu_object_header *hdr)
 518 {
 519         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 520                    hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
 521                    PFID(&hdr->loh_fid),
 522                    hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 523                    list_empty((struct list_head *)&hdr->loh_lru) ? \
 524                    "" : " lru",
 525                    hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
 526 }
 527 EXPORT_SYMBOL(lu_object_header_print);
 528
 529 /**
 530  * Print human readable representation of the \a o to the \a printer.
 531  */
 532 void lu_object_print(const struct lu_env *env, void *cookie,
 533                      lu_printer_t printer, const struct lu_object *o)
 534 {
 535         static const char ruler[] = "........................................";
 536         struct lu_object_header *top;
 537         int depth = 4;
 538
 539         top = o->lo_header;
 540         lu_object_header_print(env, cookie, printer, top);
 541         (*printer)(env, cookie, "{\n");
 542
 543         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 544                 /*
 545                  * print `.' \a depth times followed by type name and address
 546                  */
 547                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 548                            o->lo_dev->ld_type->ldt_name, o);
 549
 550                 if (o->lo_ops->loo_object_print != NULL)
 551                         (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
 552
 553                 (*printer)(env, cookie, "\n");
 554         }
 555
 556         (*printer)(env, cookie, "} header@%p\n", top);
 557 }
 558 EXPORT_SYMBOL(lu_object_print);
 559
 560 /**
 561  * Check object consistency.
 562  */
 563 int lu_object_invariant(const struct lu_object *o)
 564 {
 565         struct lu_object_header *top;
 566
 567         top = o->lo_header;
 568         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 569                 if (o->lo_ops->loo_object_invariant != NULL &&
 570                     !o->lo_ops->loo_object_invariant(o))
 571                         return 0;
 572         }
 573         return 1;
 574 }
 575
 576 static struct lu_object *htable_lookup(struct lu_site *s,
 577                                        cfs_hash_bd_t *bd,
 578                                        const struct lu_fid *f,
 579                                        wait_queue_t *waiter,
 580                                        __u64 *version)
 581 {
 582         struct lu_site_bkt_data *bkt;
 583         struct lu_object_header *h;
 584         struct hlist_node       *hnode;
 585         __u64  ver = cfs_hash_bd_version_get(bd);
 586
 587         if (*version == ver)
 588                 return ERR_PTR(-ENOENT);
 589
 590         *version = ver;
 591         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 592         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 593          * of cfs_hash, it doesn't add refcount on object. */
 594         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 595         if (hnode == NULL) {
 596                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 597                 return ERR_PTR(-ENOENT);
 598         }
 599
 600         h = container_of0(hnode, struct lu_object_header, loh_hash);
 601         if (likely(!lu_object_is_dying(h))) {
 602                 cfs_hash_get(s->ls_obj_hash, hnode);
 603                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 604                 if (!list_empty(&h->loh_lru)) {
 605                         list_del_init(&h->loh_lru);
 606                         bkt->lsb_lru_len--;
 607                 }
 608                 return lu_object_top(h);
 609         }
 610
 611         /*
 612          * Lookup found an object being destroyed this object cannot be
 613          * returned (to assure that references to dying objects are eventually
 614          * drained), and moreover, lookup has to wait until object is freed.
 615          */
 616
 617         if (likely(waiter != NULL)) {
 618                 init_waitqueue_entry_current(waiter);
 619                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 620                 set_current_state(TASK_UNINTERRUPTIBLE);
 621                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 622         }
 623
 624         return ERR_PTR(-EAGAIN);
 625 }
 626
 627 /**
 628  * Search cache for an object with the fid \a f. If such object is found,
 629  * return it. Otherwise, create new object, insert it into cache and return
 630  * it. In any case, additional reference is acquired on the returned object.
 631  */
 632 struct lu_object *lu_object_find(const struct lu_env *env,
 633                                  struct lu_device *dev, const struct lu_fid *f,
 634                                  const struct lu_object_conf *conf)
 635 {
 636         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 637 }
 638 EXPORT_SYMBOL(lu_object_find);
 639
 640 /*
 641  * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because
 642  * the calculation for the number of objects to reclaim is not covered by
 643  * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
 644  * This ensures that many concurrent threads will not accidentally purge
 645  * the entire cache.
 646  */
 647 static void lu_object_limit(const struct lu_env *env,
 648                             struct lu_device *dev)
 649 {
 650         __u64 size, nr;
 651
 652         if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
 653                 return;
 654
 655         size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
 656         nr = (__u64)lu_cache_nr;
 657         if (size > nr)
 658                 lu_site_purge(env, dev->ld_site,
 659                               MIN(size - nr, LU_CACHE_NR_MAX_ADJUST));
 660
 661         return;
 662 }
 663
 664 static struct lu_object *lu_object_new(const struct lu_env *env,
 665                                        struct lu_device *dev,
 666                                        const struct lu_fid *f,
 667                                        const struct lu_object_conf *conf)
 668 {
 669         struct lu_object        *o;
 670         cfs_hash_t              *hs;
 671         cfs_hash_bd_t            bd;
 672
 673         o = lu_object_alloc(env, dev, f, conf);
 674         if (unlikely(IS_ERR(o)))
 675                 return o;
 676
 677         hs = dev->ld_site->ls_obj_hash;
 678         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 679         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 680         cfs_hash_bd_unlock(hs, &bd, 1);
 681
 682         lu_object_limit(env, dev);
 683
 684         return o;
 685 }
 686
 687 /**
 688  * Core logic of lu_object_find*() functions.
 689  */
 690 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 691                                             struct lu_device *dev,
 692                                             const struct lu_fid *f,
 693                                             const struct lu_object_conf *conf,
 694                                             wait_queue_t *waiter)
 695 {
 696         struct lu_object      *o;
 697         struct lu_object      *shadow;
 698         struct lu_site        *s;
 699         cfs_hash_t            *hs;
 700         cfs_hash_bd_t          bd;
 701         __u64                  version = 0;
 702
 703         /*
 704          * This uses standard index maintenance protocol:
 705          *
 706          *     - search index under lock, and return object if found;
 707          *     - otherwise, unlock index, allocate new object;
 708          *     - lock index and search again;
 709          *     - if nothing is found (usual case), insert newly created
 710          *       object into index;
 711          *     - otherwise (race: other thread inserted object), free
 712          *       object just allocated.
 713          *     - unlock index;
 714          *     - return object.
 715          *
 716          * For "LOC_F_NEW" case, we are sure the object is new established.
 717          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 718          * just alloc and insert directly.
 719          *
 720          * If dying object is found during index search, add @waiter to the
 721          * site wait-queue and return ERR_PTR(-EAGAIN).
 722          */
 723         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 724                 return lu_object_new(env, dev, f, conf);
 725
 726         s  = dev->ld_site;
 727         hs = s->ls_obj_hash;
 728         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 729         o = htable_lookup(s, &bd, f, waiter, &version);
 730         cfs_hash_bd_unlock(hs, &bd, 1);
 731         if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
 732                 return o;
 733
 734         /*
 735          * Allocate new object. This may result in rather complicated
 736          * operations, including fld queries, inode loading, etc.
 737          */
 738         o = lu_object_alloc(env, dev, f, conf);
 739         if (unlikely(IS_ERR(o)))
 740                 return o;
 741
 742         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 743
 744         cfs_hash_bd_lock(hs, &bd, 1);
 745
 746         shadow = htable_lookup(s, &bd, f, waiter, &version);
 747         if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
 748                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 749                 cfs_hash_bd_unlock(hs, &bd, 1);
 750
 751                 lu_object_limit(env, dev);
 752
 753                 return o;
 754         }
 755
 756         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 757         cfs_hash_bd_unlock(hs, &bd, 1);
 758         lu_object_free(env, o);
 759         return shadow;
 760 }
 761
 762 /**
 763  * Much like lu_object_find(), but top level device of object is specifically
 764  * \a dev rather than top level device of the site. This interface allows
 765  * objects of different "stacking" to be created within the same site.
 766  */
 767 struct lu_object *lu_object_find_at(const struct lu_env *env,
 768                                     struct lu_device *dev,
 769                                     const struct lu_fid *f,
 770                                     const struct lu_object_conf *conf)
 771 {
 772         struct lu_site_bkt_data *bkt;
 773         struct lu_object        *obj;
 774         wait_queue_t           wait;
 775
 776         if (conf != NULL && conf->loc_flags & LOC_F_NOWAIT)
 777                 return lu_object_find_try(env, dev, f, conf, NULL);
 778
 779         while (1) {
 780                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 781                 if (obj != ERR_PTR(-EAGAIN))
 782                         return obj;
 783                 /*
 784                  * lu_object_find_try() already added waiter into the
 785                  * wait queue.
 786                  */
 787                 waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
 788                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 789                 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
 790         }
 791 }
 792 EXPORT_SYMBOL(lu_object_find_at);
 793
 794 /**
 795  * Find object with given fid, and return its slice belonging to given device.
 796  */
 797 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 798                                        struct lu_device *dev,
 799                                        const struct lu_fid *f,
 800                                        const struct lu_object_conf *conf)
 801 {
 802         struct lu_object *top;
 803         struct lu_object *obj;
 804
 805         top = lu_object_find(env, dev, f, conf);
 806         if (IS_ERR(top))
 807                 return top;
 808
 809         obj = lu_object_locate(top->lo_header, dev->ld_type);
 810         if (unlikely(obj == NULL)) {
 811                 lu_object_put(env, top);
 812                 obj = ERR_PTR(-ENOENT);
 813         }
 814
 815         return obj;
 816 }
 817 EXPORT_SYMBOL(lu_object_find_slice);
 818
 819 /**
 820  * Global list of all device types.
 821  */
 822 static struct list_head lu_device_types;
 823
 824 int lu_device_type_init(struct lu_device_type *ldt)
 825 {
 826         int result = 0;
 827
 828         atomic_set(&ldt->ldt_device_nr, 0);
 829         INIT_LIST_HEAD(&ldt->ldt_linkage);
 830         if (ldt->ldt_ops->ldto_init)
 831                 result = ldt->ldt_ops->ldto_init(ldt);
 832
 833         if (result == 0) {
 834                 spin_lock(&obd_types_lock);
 835                 list_add(&ldt->ldt_linkage, &lu_device_types);
 836                 spin_unlock(&obd_types_lock);
 837         }
 838
 839         return result;
 840 }
 841 EXPORT_SYMBOL(lu_device_type_init);
 842
 843 void lu_device_type_fini(struct lu_device_type *ldt)
 844 {
 845         spin_lock(&obd_types_lock);
 846         list_del_init(&ldt->ldt_linkage);
 847         spin_unlock(&obd_types_lock);
 848         if (ldt->ldt_ops->ldto_fini)
 849                 ldt->ldt_ops->ldto_fini(ldt);
 850 }
 851 EXPORT_SYMBOL(lu_device_type_fini);
 852
 853 /**
 854  * Global list of all sites on this node
 855  */
 856 static struct list_head lu_sites;
 857 static DEFINE_MUTEX(lu_sites_guard);
 858
 859 /**
 860  * Global environment used by site shrinker.
 861  */
 862 static struct lu_env lu_shrink_env;
 863
 864 struct lu_site_print_arg {
 865         struct lu_env   *lsp_env;
 866         void            *lsp_cookie;
 867         lu_printer_t     lsp_printer;
 868 };
 869
 870 static int
 871 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 872                   struct hlist_node *hnode, void *data)
 873 {
 874         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 875         struct lu_object_header  *h;
 876
 877         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 878         if (!list_empty(&h->loh_layers)) {
 879                 const struct lu_object *o;
 880
 881                 o = lu_object_top(h);
 882                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 883                                 arg->lsp_printer, o);
 884         } else {
 885                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 886                                        arg->lsp_printer, h);
 887         }
 888         return 0;
 889 }
 890
 891 /**
 892  * Print all objects in \a s.
 893  */
 894 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 895                    lu_printer_t printer)
 896 {
 897         struct lu_site_print_arg arg = {
 898                 .lsp_env     = (struct lu_env *)env,
 899                 .lsp_cookie  = cookie,
 900                 .lsp_printer = printer,
 901         };
 902
 903         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 904 }
 905 EXPORT_SYMBOL(lu_site_print);
 906
 907 /**
 908  * Return desired hash table order.
 909  */
 910 static unsigned long lu_htable_order(struct lu_device *top)
 911 {
 912         unsigned long cache_size;
 913         unsigned long bits;
 914
 915         /*
 916          * For ZFS based OSDs the cache should be disabled by default.  This
 917          * allows the ZFS ARC maximum flexibility in determining what buffers
 918          * to cache.  If Lustre has objects or buffer which it wants to ensure
 919          * always stay cached it must maintain a hold on them.
 920          */
 921         if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
 922                 lu_cache_percent = 1;
 923                 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
 924                 return LU_SITE_BITS_MIN;
 925         }
 926
 927         /*
 928          * Calculate hash table size, assuming that we want reasonable
 929          * performance when 20% of total memory is occupied by cache of
 930          * lu_objects.
 931          *
 932          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 933          */
 934         cache_size = totalram_pages;
 935
 936 #if BITS_PER_LONG == 32
 937         /* limit hashtable size for lowmem systems to low RAM */
 938         if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
 939                 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
 940 #endif
 941
 942         /* clear off unreasonable cache setting. */
 943         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 944                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 945                       " the range of (0, %u]. Will use default value: %u.\n",
 946                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 947                       LU_CACHE_PERCENT_DEFAULT);
 948
 949                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 950         }
 951         cache_size = cache_size / 100 * lu_cache_percent *
 952                 (PAGE_CACHE_SIZE / 1024);
 953
 954         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 955                 ;
 956         }
 957         return bits;
 958 }
 959
 960 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
 961                                 const void *key, unsigned mask)
 962 {
 963         struct lu_fid  *fid = (struct lu_fid *)key;
 964         __u32           hash;
 965
 966         hash = fid_flatten32(fid);
 967         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 968         hash = hash_long(hash, hs->hs_bkt_bits);
 969
 970         /* give me another random factor */
 971         hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 972
 973         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 974         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 975
 976         return hash & mask;
 977 }
 978
 979 static void *lu_obj_hop_object(struct hlist_node *hnode)
 980 {
 981         return hlist_entry(hnode, struct lu_object_header, loh_hash);
 982 }
 983
 984 static void *lu_obj_hop_key(struct hlist_node *hnode)
 985 {
 986         struct lu_object_header *h;
 987
 988         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 989         return &h->loh_fid;
 990 }
 991
 992 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
 993 {
 994         struct lu_object_header *h;
 995
 996         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 997         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 998 }
 999
1000 static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
1001 {
1002         struct lu_object_header *h;
1003
1004         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1005         atomic_inc(&h->loh_ref);
1006 }
1007
1008 static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
1009 {
1010         LBUG(); /* we should never called it */
1011 }
1012
1013 static cfs_hash_ops_t lu_site_hash_ops = {
1014         .hs_hash        = lu_obj_hop_hash,
1015         .hs_key         = lu_obj_hop_key,
1016         .hs_keycmp      = lu_obj_hop_keycmp,
1017         .hs_object      = lu_obj_hop_object,
1018         .hs_get         = lu_obj_hop_get,
1019         .hs_put_locked  = lu_obj_hop_put_locked,
1020 };
1021
1022 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1023 {
1024         spin_lock(&s->ls_ld_lock);
1025         if (list_empty(&d->ld_linkage))
1026                 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1027         spin_unlock(&s->ls_ld_lock);
1028 }
1029 EXPORT_SYMBOL(lu_dev_add_linkage);
1030
1031 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1032 {
1033         spin_lock(&s->ls_ld_lock);
1034         list_del_init(&d->ld_linkage);
1035         spin_unlock(&s->ls_ld_lock);
1036 }
1037 EXPORT_SYMBOL(lu_dev_del_linkage);
1038
1039 /**
1040   * Initialize site \a s, with \a d as the top level device.
1041   */
1042 int lu_site_init(struct lu_site *s, struct lu_device *top)
1043 {
1044         struct lu_site_bkt_data *bkt;
1045         cfs_hash_bd_t bd;
1046         char name[16];
1047         unsigned long bits;
1048         unsigned int i;
1049         ENTRY;
1050
1051         memset(s, 0, sizeof *s);
1052         mutex_init(&s->ls_purge_mutex);
1053         bits = lu_htable_order(top);
1054         snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1055         for (bits = clamp_t(typeof(bits), bits,
1056                             LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
1057              bits >= LU_SITE_BITS_MIN; bits--) {
1058                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1059                                                  bits - LU_SITE_BKT_BITS,
1060                                                  sizeof(*bkt), 0, 0,
1061                                                  &lu_site_hash_ops,
1062                                                  CFS_HASH_SPIN_BKTLOCK |
1063                                                  CFS_HASH_NO_ITEMREF |
1064                                                  CFS_HASH_DEPTH |
1065                                                  CFS_HASH_ASSERT_EMPTY |
1066                                                  CFS_HASH_COUNTER);
1067                 if (s->ls_obj_hash != NULL)
1068                         break;
1069         }
1070
1071         if (s->ls_obj_hash == NULL) {
1072                 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1073                 return -ENOMEM;
1074         }
1075
1076         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1077                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1078                 INIT_LIST_HEAD(&bkt->lsb_lru);
1079                 init_waitqueue_head(&bkt->lsb_marche_funebre);
1080         }
1081
1082         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1083         if (s->ls_stats == NULL) {
1084                 cfs_hash_putref(s->ls_obj_hash);
1085                 s->ls_obj_hash = NULL;
1086                 return -ENOMEM;
1087         }
1088
1089         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1090                              0, "created", "created");
1091         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1092                              0, "cache_hit", "cache_hit");
1093         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1094                              0, "cache_miss", "cache_miss");
1095         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1096                              0, "cache_race", "cache_race");
1097         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1098                              0, "cache_death_race", "cache_death_race");
1099         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1100                              0, "lru_purged", "lru_purged");
1101
1102         INIT_LIST_HEAD(&s->ls_linkage);
1103         s->ls_top_dev = top;
1104         top->ld_site = s;
1105         lu_device_get(top);
1106         lu_ref_add(&top->ld_reference, "site-top", s);
1107
1108         INIT_LIST_HEAD(&s->ls_ld_linkage);
1109         spin_lock_init(&s->ls_ld_lock);
1110
1111         lu_dev_add_linkage(s, top);
1112
1113         RETURN(0);
1114 }
1115 EXPORT_SYMBOL(lu_site_init);
1116
1117 /**
1118  * Finalize \a s and release its resources.
1119  */
1120 void lu_site_fini(struct lu_site *s)
1121 {
1122         mutex_lock(&lu_sites_guard);
1123         list_del_init(&s->ls_linkage);
1124         mutex_unlock(&lu_sites_guard);
1125
1126         if (s->ls_obj_hash != NULL) {
1127                 cfs_hash_putref(s->ls_obj_hash);
1128                 s->ls_obj_hash = NULL;
1129         }
1130
1131         if (s->ls_top_dev != NULL) {
1132                 s->ls_top_dev->ld_site = NULL;
1133                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1134                 lu_device_put(s->ls_top_dev);
1135                 s->ls_top_dev = NULL;
1136         }
1137
1138         if (s->ls_stats != NULL)
1139                 lprocfs_free_stats(&s->ls_stats);
1140 }
1141 EXPORT_SYMBOL(lu_site_fini);
1142
1143 /**
1144  * Called when initialization of stack for this site is completed.
1145  */
1146 int lu_site_init_finish(struct lu_site *s)
1147 {
1148         int result;
1149         mutex_lock(&lu_sites_guard);
1150         result = lu_context_refill(&lu_shrink_env.le_ctx);
1151         if (result == 0)
1152                 list_add(&s->ls_linkage, &lu_sites);
1153         mutex_unlock(&lu_sites_guard);
1154         return result;
1155 }
1156 EXPORT_SYMBOL(lu_site_init_finish);
1157
1158 /**
1159  * Acquire additional reference on device \a d
1160  */
1161 void lu_device_get(struct lu_device *d)
1162 {
1163         atomic_inc(&d->ld_ref);
1164 }
1165 EXPORT_SYMBOL(lu_device_get);
1166
1167 /**
1168  * Release reference on device \a d.
1169  */
1170 void lu_device_put(struct lu_device *d)
1171 {
1172         LASSERT(atomic_read(&d->ld_ref) > 0);
1173         atomic_dec(&d->ld_ref);
1174 }
1175 EXPORT_SYMBOL(lu_device_put);
1176
1177 /**
1178  * Initialize device \a d of type \a t.
1179  */
1180 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1181 {
1182         if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1183             t->ldt_ops->ldto_start != NULL)
1184                 t->ldt_ops->ldto_start(t);
1185
1186         memset(d, 0, sizeof *d);
1187         d->ld_type = t;
1188         lu_ref_init(&d->ld_reference);
1189         INIT_LIST_HEAD(&d->ld_linkage);
1190
1191         return 0;
1192 }
1193 EXPORT_SYMBOL(lu_device_init);
1194
1195 /**
1196  * Finalize device \a d.
1197  */
1198 void lu_device_fini(struct lu_device *d)
1199 {
1200         struct lu_device_type *t = d->ld_type;
1201
1202         if (d->ld_obd != NULL) {
1203                 d->ld_obd->obd_lu_dev = NULL;
1204                 d->ld_obd = NULL;
1205         }
1206
1207         lu_ref_fini(&d->ld_reference);
1208         LASSERTF(atomic_read(&d->ld_ref) == 0,
1209                  "Refcount is %u\n", atomic_read(&d->ld_ref));
1210         LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1211
1212         if (atomic_dec_and_test(&t->ldt_device_nr) &&
1213             t->ldt_ops->ldto_stop != NULL)
1214                 t->ldt_ops->ldto_stop(t);
1215 }
1216 EXPORT_SYMBOL(lu_device_fini);
1217
1218 /**
1219  * Initialize object \a o that is part of compound object \a h and was created
1220  * by device \a d.
1221  */
1222 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1223                    struct lu_device *d)
1224 {
1225         memset(o, 0, sizeof(*o));
1226         o->lo_header = h;
1227         o->lo_dev = d;
1228         lu_device_get(d);
1229         lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1230         INIT_LIST_HEAD(&o->lo_linkage);
1231
1232         return 0;
1233 }
1234 EXPORT_SYMBOL(lu_object_init);
1235
1236 /**
1237  * Finalize object and release its resources.
1238  */
1239 void lu_object_fini(struct lu_object *o)
1240 {
1241         struct lu_device *dev = o->lo_dev;
1242
1243         LASSERT(list_empty(&o->lo_linkage));
1244
1245         if (dev != NULL) {
1246                 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1247                               "lu_object", o);
1248                 lu_device_put(dev);
1249                 o->lo_dev = NULL;
1250         }
1251 }
1252 EXPORT_SYMBOL(lu_object_fini);
1253
1254 /**
1255  * Add object \a o as first layer of compound object \a h
1256  *
1257  * This is typically called by the ->ldo_object_alloc() method of top-level
1258  * device.
1259  */
1260 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1261 {
1262         list_move(&o->lo_linkage, &h->loh_layers);
1263 }
1264 EXPORT_SYMBOL(lu_object_add_top);
1265
1266 /**
1267  * Add object \a o as a layer of compound object, going after \a before.
1268  *
1269  * This is typically called by the ->ldo_object_alloc() method of \a
1270  * before->lo_dev.
1271  */
1272 void lu_object_add(struct lu_object *before, struct lu_object *o)
1273 {
1274         list_move(&o->lo_linkage, &before->lo_linkage);
1275 }
1276 EXPORT_SYMBOL(lu_object_add);
1277
1278 /**
1279  * Initialize compound object.
1280  */
1281 int lu_object_header_init(struct lu_object_header *h)
1282 {
1283         memset(h, 0, sizeof *h);
1284         atomic_set(&h->loh_ref, 1);
1285         INIT_HLIST_NODE(&h->loh_hash);
1286         INIT_LIST_HEAD(&h->loh_lru);
1287         INIT_LIST_HEAD(&h->loh_layers);
1288         lu_ref_init(&h->loh_reference);
1289         return 0;
1290 }
1291 EXPORT_SYMBOL(lu_object_header_init);
1292
1293 /**
1294  * Finalize compound object.
1295  */
1296 void lu_object_header_fini(struct lu_object_header *h)
1297 {
1298         LASSERT(list_empty(&h->loh_layers));
1299         LASSERT(list_empty(&h->loh_lru));
1300         LASSERT(hlist_unhashed(&h->loh_hash));
1301         lu_ref_fini(&h->loh_reference);
1302 }
1303 EXPORT_SYMBOL(lu_object_header_fini);
1304
1305 /**
1306  * Given a compound object, find its slice, corresponding to the device type
1307  * \a dtype.
1308  */
1309 struct lu_object *lu_object_locate(struct lu_object_header *h,
1310                                    const struct lu_device_type *dtype)
1311 {
1312         struct lu_object *o;
1313
1314         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1315                 if (o->lo_dev->ld_type == dtype)
1316                         return o;
1317         }
1318         return NULL;
1319 }
1320 EXPORT_SYMBOL(lu_object_locate);
1321
1322 /**
1323  * Finalize and free devices in the device stack.
1324  *
1325  * Finalize device stack by purging object cache, and calling
1326  * lu_device_type_operations::ldto_device_fini() and
1327  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1328  */
1329 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1330 {
1331         struct lu_site   *site = top->ld_site;
1332         struct lu_device *scan;
1333         struct lu_device *next;
1334
1335         lu_site_purge(env, site, ~0);
1336         for (scan = top; scan != NULL; scan = next) {
1337                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1338                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1339                 lu_device_put(scan);
1340         }
1341
1342         /* purge again. */
1343         lu_site_purge(env, site, ~0);
1344
1345         for (scan = top; scan != NULL; scan = next) {
1346                 const struct lu_device_type *ldt = scan->ld_type;
1347                 struct obd_type             *type;
1348
1349                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1350                 type = ldt->ldt_obd_type;
1351                 if (type != NULL) {
1352                         type->typ_refcnt--;
1353                         class_put_type(type);
1354                 }
1355         }
1356 }
1357
1358 enum {
1359         /**
1360          * Maximal number of tld slots.
1361          */
1362         LU_CONTEXT_KEY_NR = 40
1363 };
1364
1365 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1366
1367 static DEFINE_SPINLOCK(lu_keys_guard);
1368
1369 /**
1370  * Global counter incremented whenever key is registered, unregistered,
1371  * revived or quiesced. This is used to void unnecessary calls to
1372  * lu_context_refill(). No locking is provided, as initialization and shutdown
1373  * are supposed to be externally serialized.
1374  */
1375 static unsigned key_set_version = 0;
1376
1377 /**
1378  * Register new key.
1379  */
1380 int lu_context_key_register(struct lu_context_key *key)
1381 {
1382         int result;
1383         unsigned int i;
1384
1385         LASSERT(key->lct_init != NULL);
1386         LASSERT(key->lct_fini != NULL);
1387         LASSERT(key->lct_tags != 0);
1388         LASSERT(key->lct_owner != NULL);
1389
1390         result = -ENFILE;
1391         spin_lock(&lu_keys_guard);
1392         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1393                 if (lu_keys[i] == NULL) {
1394                         key->lct_index = i;
1395                         atomic_set(&key->lct_used, 1);
1396                         lu_keys[i] = key;
1397                         lu_ref_init(&key->lct_reference);
1398                         result = 0;
1399                         ++key_set_version;
1400                         break;
1401                 }
1402         }
1403         spin_unlock(&lu_keys_guard);
1404         return result;
1405 }
1406 EXPORT_SYMBOL(lu_context_key_register);
1407
1408 static void key_fini(struct lu_context *ctx, int index)
1409 {
1410         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1411                 struct lu_context_key *key;
1412
1413                 key = lu_keys[index];
1414                 LASSERT(key != NULL);
1415                 LASSERT(key->lct_fini != NULL);
1416                 LASSERT(atomic_read(&key->lct_used) > 1);
1417
1418                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1419                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1420                 atomic_dec(&key->lct_used);
1421
1422                 LASSERT(key->lct_owner != NULL);
1423                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1424                         LINVRNT(module_refcount(key->lct_owner) > 0);
1425                         module_put(key->lct_owner);
1426                 }
1427                 ctx->lc_value[index] = NULL;
1428         }
1429 }
1430
1431 /**
1432  * Deregister key.
1433  */
1434 void lu_context_key_degister(struct lu_context_key *key)
1435 {
1436         LASSERT(atomic_read(&key->lct_used) >= 1);
1437         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1438
1439         lu_context_key_quiesce(key);
1440
1441         ++key_set_version;
1442         spin_lock(&lu_keys_guard);
1443         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1444         if (lu_keys[key->lct_index]) {
1445                 lu_keys[key->lct_index] = NULL;
1446                 lu_ref_fini(&key->lct_reference);
1447         }
1448         spin_unlock(&lu_keys_guard);
1449
1450         LASSERTF(atomic_read(&key->lct_used) == 1,
1451                  "key has instances: %d\n",
1452                  atomic_read(&key->lct_used));
1453 }
1454 EXPORT_SYMBOL(lu_context_key_degister);
1455
1456 /**
1457  * Register a number of keys. This has to be called after all keys have been
1458  * initialized by a call to LU_CONTEXT_KEY_INIT().
1459  */
1460 int lu_context_key_register_many(struct lu_context_key *k, ...)
1461 {
1462         struct lu_context_key *key = k;
1463         va_list args;
1464         int result;
1465
1466         va_start(args, k);
1467         do {
1468                 result = lu_context_key_register(key);
1469                 if (result)
1470                         break;
1471                 key = va_arg(args, struct lu_context_key *);
1472         } while (key != NULL);
1473         va_end(args);
1474
1475         if (result != 0) {
1476                 va_start(args, k);
1477                 while (k != key) {
1478                         lu_context_key_degister(k);
1479                         k = va_arg(args, struct lu_context_key *);
1480                 }
1481                 va_end(args);
1482         }
1483
1484         return result;
1485 }
1486 EXPORT_SYMBOL(lu_context_key_register_many);
1487
1488 /**
1489  * De-register a number of keys. This is a dual to
1490  * lu_context_key_register_many().
1491  */
1492 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1493 {
1494         va_list args;
1495
1496         va_start(args, k);
1497         do {
1498                 lu_context_key_degister(k);
1499                 k = va_arg(args, struct lu_context_key*);
1500         } while (k != NULL);
1501         va_end(args);
1502 }
1503 EXPORT_SYMBOL(lu_context_key_degister_many);
1504
1505 /**
1506  * Revive a number of keys.
1507  */
1508 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1509 {
1510         va_list args;
1511
1512         va_start(args, k);
1513         do {
1514                 lu_context_key_revive(k);
1515                 k = va_arg(args, struct lu_context_key*);
1516         } while (k != NULL);
1517         va_end(args);
1518 }
1519 EXPORT_SYMBOL(lu_context_key_revive_many);
1520
1521 /**
1522  * Quiescent a number of keys.
1523  */
1524 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1525 {
1526         va_list args;
1527
1528         va_start(args, k);
1529         do {
1530                 lu_context_key_quiesce(k);
1531                 k = va_arg(args, struct lu_context_key*);
1532         } while (k != NULL);
1533         va_end(args);
1534 }
1535 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1536
1537 /**
1538  * Return value associated with key \a key in context \a ctx.
1539  */
1540 void *lu_context_key_get(const struct lu_context *ctx,
1541                          const struct lu_context_key *key)
1542 {
1543         LINVRNT(ctx->lc_state == LCS_ENTERED);
1544         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1545         LASSERT(lu_keys[key->lct_index] == key);
1546         return ctx->lc_value[key->lct_index];
1547 }
1548 EXPORT_SYMBOL(lu_context_key_get);
1549
1550 /**
1551  * List of remembered contexts. XXX document me.
1552  */
1553 static struct list_head lu_context_remembered;
1554
1555 /**
1556  * Destroy \a key in all remembered contexts. This is used to destroy key
1557  * values in "shared" contexts (like service threads), when a module owning
1558  * the key is about to be unloaded.
1559  */
1560 void lu_context_key_quiesce(struct lu_context_key *key)
1561 {
1562         struct lu_context *ctx;
1563         extern unsigned cl_env_cache_purge(unsigned nr);
1564
1565         if (!(key->lct_tags & LCT_QUIESCENT)) {
1566                 /*
1567                  * XXX layering violation.
1568                  */
1569                 cl_env_cache_purge(~0);
1570                 key->lct_tags |= LCT_QUIESCENT;
1571                 /*
1572                  * XXX memory barrier has to go here.
1573                  */
1574                 spin_lock(&lu_keys_guard);
1575                 list_for_each_entry(ctx, &lu_context_remembered,
1576                                     lc_remember)
1577                         key_fini(ctx, key->lct_index);
1578                 spin_unlock(&lu_keys_guard);
1579                 ++key_set_version;
1580         }
1581 }
1582
1583 void lu_context_key_revive(struct lu_context_key *key)
1584 {
1585         key->lct_tags &= ~LCT_QUIESCENT;
1586         ++key_set_version;
1587 }
1588
1589 static void keys_fini(struct lu_context *ctx)
1590 {
1591         unsigned int i;
1592
1593         if (ctx->lc_value == NULL)
1594                 return;
1595
1596         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1597                 key_fini(ctx, i);
1598
1599         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1600         ctx->lc_value = NULL;
1601 }
1602
1603 static int keys_fill(struct lu_context *ctx)
1604 {
1605         unsigned int i;
1606
1607         LINVRNT(ctx->lc_value != NULL);
1608         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1609                 struct lu_context_key *key;
1610
1611                 key = lu_keys[i];
1612                 if (ctx->lc_value[i] == NULL && key != NULL &&
1613                     (key->lct_tags & ctx->lc_tags) &&
1614                     /*
1615                      * Don't create values for a LCT_QUIESCENT key, as this
1616                      * will pin module owning a key.
1617                      */
1618                     !(key->lct_tags & LCT_QUIESCENT)) {
1619                         void *value;
1620
1621                         LINVRNT(key->lct_init != NULL);
1622                         LINVRNT(key->lct_index == i);
1623
1624                         value = key->lct_init(ctx, key);
1625                         if (unlikely(IS_ERR(value)))
1626                                 return PTR_ERR(value);
1627
1628                         LASSERT(key->lct_owner != NULL);
1629                         if (!(ctx->lc_tags & LCT_NOREF))
1630                                 try_module_get(key->lct_owner);
1631                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1632                         atomic_inc(&key->lct_used);
1633                         /*
1634                          * This is the only place in the code, where an
1635                          * element of ctx->lc_value[] array is set to non-NULL
1636                          * value.
1637                          */
1638                         ctx->lc_value[i] = value;
1639                         if (key->lct_exit != NULL)
1640                                 ctx->lc_tags |= LCT_HAS_EXIT;
1641                 }
1642                 ctx->lc_version = key_set_version;
1643         }
1644         return 0;
1645 }
1646
1647 static int keys_init(struct lu_context *ctx)
1648 {
1649         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1650         if (likely(ctx->lc_value != NULL))
1651                 return keys_fill(ctx);
1652
1653         return -ENOMEM;
1654 }
1655
1656 /**
1657  * Initialize context data-structure. Create values for all keys.
1658  */
1659 int lu_context_init(struct lu_context *ctx, __u32 tags)
1660 {
1661         int     rc;
1662
1663         memset(ctx, 0, sizeof *ctx);
1664         ctx->lc_state = LCS_INITIALIZED;
1665         ctx->lc_tags = tags;
1666         if (tags & LCT_REMEMBER) {
1667                 spin_lock(&lu_keys_guard);
1668                 list_add(&ctx->lc_remember, &lu_context_remembered);
1669                 spin_unlock(&lu_keys_guard);
1670         } else {
1671                 INIT_LIST_HEAD(&ctx->lc_remember);
1672         }
1673
1674         rc = keys_init(ctx);
1675         if (rc != 0)
1676                 lu_context_fini(ctx);
1677
1678         return rc;
1679 }
1680 EXPORT_SYMBOL(lu_context_init);
1681
1682 /**
1683  * Finalize context data-structure. Destroy key values.
1684  */
1685 void lu_context_fini(struct lu_context *ctx)
1686 {
1687         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1688         ctx->lc_state = LCS_FINALIZED;
1689
1690         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1691                 LASSERT(list_empty(&ctx->lc_remember));
1692                 keys_fini(ctx);
1693
1694         } else { /* could race with key degister */
1695                 spin_lock(&lu_keys_guard);
1696                 keys_fini(ctx);
1697                 list_del_init(&ctx->lc_remember);
1698                 spin_unlock(&lu_keys_guard);
1699         }
1700 }
1701 EXPORT_SYMBOL(lu_context_fini);
1702
1703 /**
1704  * Called before entering context.
1705  */
1706 void lu_context_enter(struct lu_context *ctx)
1707 {
1708         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1709         ctx->lc_state = LCS_ENTERED;
1710 }
1711 EXPORT_SYMBOL(lu_context_enter);
1712
1713 /**
1714  * Called after exiting from \a ctx
1715  */
1716 void lu_context_exit(struct lu_context *ctx)
1717 {
1718         unsigned int i;
1719
1720         LINVRNT(ctx->lc_state == LCS_ENTERED);
1721         ctx->lc_state = LCS_LEFT;
1722         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1723                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1724                         if (ctx->lc_value[i] != NULL) {
1725                                 struct lu_context_key *key;
1726
1727                                 key = lu_keys[i];
1728                                 LASSERT(key != NULL);
1729                                 if (key->lct_exit != NULL)
1730                                         key->lct_exit(ctx,
1731                                                       key, ctx->lc_value[i]);
1732                         }
1733                 }
1734         }
1735 }
1736 EXPORT_SYMBOL(lu_context_exit);
1737
1738 /**
1739  * Allocate for context all missing keys that were registered after context
1740  * creation. key_set_version is only changed in rare cases when modules
1741  * are loaded and removed.
1742  */
1743 int lu_context_refill(struct lu_context *ctx)
1744 {
1745         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1746 }
1747
1748 /**
1749  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1750  * obd being added. Currently, this is only used on client side, specifically
1751  * for echo device client, for other stack (like ptlrpc threads), context are
1752  * predefined when the lu_device type are registered, during the module probe
1753  * phase.
1754  */
1755 __u32 lu_context_tags_default = 0;
1756 __u32 lu_session_tags_default = 0;
1757
1758 void lu_context_tags_update(__u32 tags)
1759 {
1760         spin_lock(&lu_keys_guard);
1761         lu_context_tags_default |= tags;
1762         key_set_version++;
1763         spin_unlock(&lu_keys_guard);
1764 }
1765 EXPORT_SYMBOL(lu_context_tags_update);
1766
1767 void lu_context_tags_clear(__u32 tags)
1768 {
1769         spin_lock(&lu_keys_guard);
1770         lu_context_tags_default &= ~tags;
1771         key_set_version++;
1772         spin_unlock(&lu_keys_guard);
1773 }
1774 EXPORT_SYMBOL(lu_context_tags_clear);
1775
1776 void lu_session_tags_update(__u32 tags)
1777 {
1778         spin_lock(&lu_keys_guard);
1779         lu_session_tags_default |= tags;
1780         key_set_version++;
1781         spin_unlock(&lu_keys_guard);
1782 }
1783 EXPORT_SYMBOL(lu_session_tags_update);
1784
1785 void lu_session_tags_clear(__u32 tags)
1786 {
1787         spin_lock(&lu_keys_guard);
1788         lu_session_tags_default &= ~tags;
1789         key_set_version++;
1790         spin_unlock(&lu_keys_guard);
1791 }
1792 EXPORT_SYMBOL(lu_session_tags_clear);
1793
1794 int lu_env_init(struct lu_env *env, __u32 tags)
1795 {
1796         int result;
1797
1798         env->le_ses = NULL;
1799         result = lu_context_init(&env->le_ctx, tags);
1800         if (likely(result == 0))
1801                 lu_context_enter(&env->le_ctx);
1802         return result;
1803 }
1804 EXPORT_SYMBOL(lu_env_init);
1805
1806 void lu_env_fini(struct lu_env *env)
1807 {
1808         lu_context_exit(&env->le_ctx);
1809         lu_context_fini(&env->le_ctx);
1810         env->le_ses = NULL;
1811 }
1812 EXPORT_SYMBOL(lu_env_fini);
1813
1814 int lu_env_refill(struct lu_env *env)
1815 {
1816         int result;
1817
1818         result = lu_context_refill(&env->le_ctx);
1819         if (result == 0 && env->le_ses != NULL)
1820                 result = lu_context_refill(env->le_ses);
1821         return result;
1822 }
1823 EXPORT_SYMBOL(lu_env_refill);
1824
1825 /**
1826  * Currently, this API will only be used by echo client.
1827  * Because echo client and normal lustre client will share
1828  * same cl_env cache. So echo client needs to refresh
1829  * the env context after it get one from the cache, especially
1830  * when normal client and echo client co-exist in the same client.
1831  */
1832 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1833                           __u32 stags)
1834 {
1835         int    result;
1836
1837         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1838                 env->le_ctx.lc_version = 0;
1839                 env->le_ctx.lc_tags |= ctags;
1840         }
1841
1842         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1843                 env->le_ses->lc_version = 0;
1844                 env->le_ses->lc_tags |= stags;
1845         }
1846
1847         result = lu_env_refill(env);
1848
1849         return result;
1850 }
1851 EXPORT_SYMBOL(lu_env_refill_by_tags);
1852
1853 static struct shrinker *lu_site_shrinker;
1854
1855 typedef struct lu_site_stats{
1856         unsigned        lss_populated;
1857         unsigned        lss_max_search;
1858         unsigned        lss_total;
1859         unsigned        lss_busy;
1860 } lu_site_stats_t;
1861
1862 static void lu_site_stats_get(cfs_hash_t *hs,
1863                               lu_site_stats_t *stats, int populated)
1864 {
1865         cfs_hash_bd_t bd;
1866         unsigned int  i;
1867
1868         cfs_hash_for_each_bucket(hs, &bd, i) {
1869                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1870                 struct hlist_head       *hhead;
1871
1872                 cfs_hash_bd_lock(hs, &bd, 1);
1873                 stats->lss_busy  +=
1874                         cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len;
1875                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1876                 stats->lss_max_search = max((int)stats->lss_max_search,
1877                                             cfs_hash_bd_depmax_get(&bd));
1878                 if (!populated) {
1879                         cfs_hash_bd_unlock(hs, &bd, 1);
1880                         continue;
1881                 }
1882
1883                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1884                         if (!hlist_empty(hhead))
1885                                 stats->lss_populated++;
1886                 }
1887                 cfs_hash_bd_unlock(hs, &bd, 1);
1888         }
1889 }
1890
1891
1892 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1893                                            struct shrink_control *sc)
1894 {
1895         lu_site_stats_t stats;
1896         struct lu_site *s;
1897         struct lu_site *tmp;
1898         unsigned long cached = 0;
1899
1900         if (!(sc->gfp_mask & __GFP_FS))
1901                 return 0;
1902
1903         mutex_lock(&lu_sites_guard);
1904         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1905                 memset(&stats, 0, sizeof(stats));
1906                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1907                 cached += stats.lss_total - stats.lss_busy;
1908         }
1909         mutex_unlock(&lu_sites_guard);
1910
1911         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1912         CDEBUG(D_INODE, "%ld objects cached\n", cached);
1913         return cached;
1914 }
1915
1916 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1917                                           struct shrink_control *sc)
1918 {
1919         struct lu_site *s;
1920         struct lu_site *tmp;
1921         unsigned long remain = sc->nr_to_scan;
1922         LIST_HEAD(splice);
1923
1924         if (!(sc->gfp_mask & __GFP_FS))
1925                 /* We must not take the lu_sites_guard lock when
1926                  * __GFP_FS is *not* set because of the deadlock
1927                  * possibility detailed above. Additionally,
1928                  * since we cannot determine the number of
1929                  * objects in the cache without taking this
1930                  * lock, we're in a particularly tough spot. As
1931                  * a result, we'll just lie and say our cache is
1932                  * empty. This _should_ be ok, as we can't
1933                  * reclaim objects when __GFP_FS is *not* set
1934                  * anyways.
1935                  */
1936                 return SHRINK_STOP;
1937
1938         mutex_lock(&lu_sites_guard);
1939         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1940                 remain = lu_site_purge(&lu_shrink_env, s, remain);
1941                 /*
1942                  * Move just shrunk site to the tail of site list to
1943                  * assure shrinking fairness.
1944                  */
1945                 list_move_tail(&s->ls_linkage, &splice);
1946         }
1947         list_splice(&splice, lu_sites.prev);
1948         mutex_unlock(&lu_sites_guard);
1949
1950         return sc->nr_to_scan - remain;
1951 }
1952
1953 #ifndef HAVE_SHRINKER_COUNT
1954 /*
1955  * There exists a potential lock inversion deadlock scenario when using
1956  * Lustre on top of ZFS. This occurs between one of ZFS's
1957  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
1958  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
1959  * while thread B will take the ht_lock and sleep on the lu_sites_guard
1960  * lock. Obviously neither thread will wake and drop their respective hold
1961  * on their lock.
1962  *
1963  * To prevent this from happening we must ensure the lu_sites_guard lock is
1964  * not taken while down this code path. ZFS reliably does not set the
1965  * __GFP_FS bit in its code paths, so this can be used to determine if it
1966  * is safe to take the lu_sites_guard lock.
1967  *
1968  * Ideally we should accurately return the remaining number of cached
1969  * objects without taking the  lu_sites_guard lock, but this is not
1970  * possible in the current implementation.
1971  */
1972 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
1973 {
1974         int cached = 0;
1975         struct shrink_control scv = {
1976                  .nr_to_scan = shrink_param(sc, nr_to_scan),
1977                  .gfp_mask   = shrink_param(sc, gfp_mask)
1978         };
1979 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
1980         struct shrinker* shrinker = NULL;
1981 #endif
1982
1983
1984         CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
1985
1986         lu_cache_shrink_scan(shrinker, &scv);
1987
1988         cached = lu_cache_shrink_count(shrinker, &scv);
1989         if (scv.nr_to_scan == 0)
1990                 CDEBUG(D_INODE, "%d objects cached\n", cached);
1991         return cached;
1992 }
1993
1994 #endif /* HAVE_SHRINKER_COUNT */
1995
1996
1997 /*
1998  * Debugging stuff.
1999  */
2000
2001 /**
2002  * Environment to be used in debugger, contains all tags.
2003  */
2004 static struct lu_env lu_debugging_env;
2005
2006 /**
2007  * Debugging printer function using printk().
2008  */
2009 int lu_printk_printer(const struct lu_env *env,
2010                       void *unused, const char *format, ...)
2011 {
2012         va_list args;
2013
2014         va_start(args, format);
2015         vprintk(format, args);
2016         va_end(args);
2017         return 0;
2018 }
2019
2020 int lu_debugging_setup(void)
2021 {
2022         return lu_env_init(&lu_debugging_env, ~0);
2023 }
2024
2025 void lu_context_keys_dump(void)
2026 {
2027         unsigned int i;
2028
2029         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2030                 struct lu_context_key *key;
2031
2032                 key = lu_keys[i];
2033                 if (key != NULL) {
2034                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2035                                i, key, key->lct_tags,
2036                                key->lct_init, key->lct_fini, key->lct_exit,
2037                                key->lct_index, atomic_read(&key->lct_used),
2038                                key->lct_owner ? key->lct_owner->name : "",
2039                                key->lct_owner);
2040                         lu_ref_print(&key->lct_reference);
2041                 }
2042         }
2043 }
2044
2045 /**
2046  * Initialization of global lu_* data.
2047  */
2048 int lu_global_init(void)
2049 {
2050         int result;
2051         DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2052                          lu_cache_shrink_count, lu_cache_shrink_scan);
2053
2054         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2055
2056         INIT_LIST_HEAD(&lu_device_types);
2057         INIT_LIST_HEAD(&lu_context_remembered);
2058         INIT_LIST_HEAD(&lu_sites);
2059
2060         result = lu_ref_global_init();
2061         if (result != 0)
2062                 return result;
2063
2064         LU_CONTEXT_KEY_INIT(&lu_global_key);
2065         result = lu_context_key_register(&lu_global_key);
2066         if (result != 0)
2067                 return result;
2068
2069         /*
2070          * At this level, we don't know what tags are needed, so allocate them
2071          * conservatively. This should not be too bad, because this
2072          * environment is global.
2073          */
2074         mutex_lock(&lu_sites_guard);
2075         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2076         mutex_unlock(&lu_sites_guard);
2077         if (result != 0)
2078                 return result;
2079
2080         /*
2081          * seeks estimation: 3 seeks to read a record from oi, one to read
2082          * inode, one for ea. Unfortunately setting this high value results in
2083          * lu_object/inode cache consuming all the memory.
2084          */
2085         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2086         if (lu_site_shrinker == NULL)
2087                 return -ENOMEM;
2088
2089         return result;
2090 }
2091
2092 /**
2093  * Dual to lu_global_init().
2094  */
2095 void lu_global_fini(void)
2096 {
2097         if (lu_site_shrinker != NULL) {
2098                 remove_shrinker(lu_site_shrinker);
2099                 lu_site_shrinker = NULL;
2100         }
2101
2102         lu_context_key_degister(&lu_global_key);
2103
2104         /*
2105          * Tear shrinker environment down _after_ de-registering
2106          * lu_global_key, because the latter has a value in the former.
2107          */
2108         mutex_lock(&lu_sites_guard);
2109         lu_env_fini(&lu_shrink_env);
2110         mutex_unlock(&lu_sites_guard);
2111
2112         lu_ref_global_fini();
2113 }
2114
2115 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2116 {
2117 #ifdef CONFIG_PROC_FS
2118         struct lprocfs_counter ret;
2119
2120         lprocfs_stats_collect(stats, idx, &ret);
2121         return (__u32)ret.lc_count;
2122 #else
2123         return 0;
2124 #endif
2125 }
2126
2127 /**
2128  * Output site statistical counters into a buffer. Suitable for
2129  * lprocfs_rd_*()-style functions.
2130  */
2131 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2132 {
2133         lu_site_stats_t stats;
2134
2135         memset(&stats, 0, sizeof(stats));
2136         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2137
2138         return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2139                           stats.lss_busy,
2140                           stats.lss_total,
2141                           stats.lss_populated,
2142                           CFS_HASH_NHLIST(s->ls_obj_hash),
2143                           stats.lss_max_search,
2144                           ls_stats_read(s->ls_stats, LU_SS_CREATED),
2145                           ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2146                           ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2147                           ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2148                           ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2149                           ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2150 }
2151 EXPORT_SYMBOL(lu_site_stats_seq_print);
2152
2153 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2154 {
2155         lu_site_stats_t stats;
2156
2157         memset(&stats, 0, sizeof(stats));
2158         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2159
2160         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2161                         stats.lss_busy,
2162                         stats.lss_total,
2163                         stats.lss_populated,
2164                         CFS_HASH_NHLIST(s->ls_obj_hash),
2165                         stats.lss_max_search,
2166                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
2167                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2168                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2169                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2170                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2171                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2172 }
2173
2174 /**
2175  * Helper function to initialize a number of kmem slab caches at once.
2176  */
2177 int lu_kmem_init(struct lu_kmem_descr *caches)
2178 {
2179         int result;
2180         struct lu_kmem_descr *iter = caches;
2181
2182         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2183                 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2184                                                      iter->ckd_size,
2185                                                      0, 0, NULL);
2186                 if (*iter->ckd_cache == NULL) {
2187                         result = -ENOMEM;
2188                         /* free all previously allocated caches */
2189                         lu_kmem_fini(caches);
2190                         break;
2191                 }
2192         }
2193         return result;
2194 }
2195 EXPORT_SYMBOL(lu_kmem_init);
2196
2197 /**
2198  * Helper function to finalize a number of kmem slab cached at once. Dual to
2199  * lu_kmem_init().
2200  */
2201 void lu_kmem_fini(struct lu_kmem_descr *caches)
2202 {
2203         for (; caches->ckd_cache != NULL; ++caches) {
2204                 if (*caches->ckd_cache != NULL) {
2205                         kmem_cache_destroy(*caches->ckd_cache);
2206                         *caches->ckd_cache = NULL;
2207                 }
2208         }
2209 }
2210 EXPORT_SYMBOL(lu_kmem_fini);
2211
2212 /**
2213  * Temporary solution to be able to assign fid in ->do_create()
2214  * till we have fully-functional OST fids
2215  */
2216 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2217                           const struct lu_fid *fid)
2218 {
2219         struct lu_site          *s = o->lo_dev->ld_site;
2220         struct lu_fid           *old = &o->lo_header->loh_fid;
2221         struct lu_object        *shadow;
2222         wait_queue_t             waiter;
2223         cfs_hash_t              *hs;
2224         cfs_hash_bd_t            bd;
2225         __u64                    version = 0;
2226
2227         LASSERT(fid_is_zero(old));
2228
2229         hs = s->ls_obj_hash;
2230         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2231         shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2232         /* supposed to be unique */
2233         LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2234         *old = *fid;
2235         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2236         cfs_hash_bd_unlock(hs, &bd, 1);
2237 }
2238 EXPORT_SYMBOL(lu_object_assign_fid);
2239
2240 /**
2241  * allocates object with 0 (non-assiged) fid
2242  * XXX: temporary solution to be able to assign fid in ->do_create()
2243  *      till we have fully-functional OST fids
2244  */
2245 struct lu_object *lu_object_anon(const struct lu_env *env,
2246                                  struct lu_device *dev,
2247                                  const struct lu_object_conf *conf)
2248 {
2249         struct lu_fid     fid;
2250         struct lu_object *o;
2251
2252         fid_zero(&fid);
2253         o = lu_object_alloc(env, dev, &fid, conf);
2254
2255         return o;
2256 }
2257 EXPORT_SYMBOL(lu_object_anon);
2258
2259 struct lu_buf LU_BUF_NULL = {
2260         .lb_buf = NULL,
2261         .lb_len = 0
2262 };
2263 EXPORT_SYMBOL(LU_BUF_NULL);
2264
2265 void lu_buf_free(struct lu_buf *buf)
2266 {
2267         LASSERT(buf);
2268         if (buf->lb_buf) {
2269                 LASSERT(buf->lb_len > 0);
2270                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2271                 buf->lb_buf = NULL;
2272                 buf->lb_len = 0;
2273         }
2274 }
2275 EXPORT_SYMBOL(lu_buf_free);
2276
2277 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2278 {
2279         LASSERT(buf);
2280         LASSERT(buf->lb_buf == NULL);
2281         LASSERT(buf->lb_len == 0);
2282         OBD_ALLOC_LARGE(buf->lb_buf, size);
2283         if (likely(buf->lb_buf))
2284                 buf->lb_len = size;
2285 }
2286 EXPORT_SYMBOL(lu_buf_alloc);
2287
2288 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2289 {
2290         lu_buf_free(buf);
2291         lu_buf_alloc(buf, size);
2292 }
2293 EXPORT_SYMBOL(lu_buf_realloc);
2294
2295 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2296 {
2297         if (buf->lb_buf == NULL && buf->lb_len == 0)
2298                 lu_buf_alloc(buf, len);
2299
2300         if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2301                 lu_buf_realloc(buf, len);
2302
2303         return buf;
2304 }
2305 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2306
2307 /**
2308  * Increase the size of the \a buf.
2309  * preserves old data in buffer
2310  * old buffer remains unchanged on error
2311  * \retval 0 or -ENOMEM
2312  */
2313 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2314 {
2315         char *ptr;
2316
2317         if (len <= buf->lb_len)
2318                 return 0;
2319
2320         OBD_ALLOC_LARGE(ptr, len);
2321         if (ptr == NULL)
2322                 return -ENOMEM;
2323
2324         /* Free the old buf */
2325         if (buf->lb_buf != NULL) {
2326                 memcpy(ptr, buf->lb_buf, buf->lb_len);
2327                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2328         }
2329
2330         buf->lb_buf = ptr;
2331         buf->lb_len = len;
2332         return 0;
2333 }